Table 1.

Summary of hyper-parameters and default settings of gcForest. Bold font highlights hyper-parameters with relatively larger influence; ‘?' indicates default value unknown, or a general requirement for different settings for different tasks.

Deep neural networks (e.g. convolutional neural networks)	gcForest
Type of activation functions:	Type of forests:
Sigmoid, ReLU, tanh, linear, etc.	Completely random forest, random forest, etc.
Architecture configurations:	Forest in multi-grained scanning:
No. hidden layers: ?	No. forests: {2}
No. nodes in hidden layer: ?	No. trees in each forest: {500}
No. feature maps: ?	Tree growth: till pure leaf, or reach depth 100
Kernel size: ?	Sliding window size: {⌊d/16⌋, ⌊d/8⌋, ⌊d/4⌋}
Optimization configurations:	Forest in cascade:
Learning rate: ?	No. forests: {8}
Dropout: {0.25/0.50}	No. trees in each forest: {500}
Momentum: ?	Tree growth: till pure leaf
L1/L2 weight regularization penalty: ?
Weight initialization: Uniform, glorot_normal, glorot_uni, etc.
Batch size: {32/64/128}